Key Takeaways
- Protein engineering faces an astronomical search space – a 50‑amino‑acid protein has ~1.13 × 10⁶⁵ possible variants, far beyond experimental feasibility.
- AI can model this space, but its effectiveness hinges on having sufficient, high‑quality experimental data for training.
- Han Xiao’s team at Rice University introduced Sequence Display, a barcoding platform that generates >10 million activity‑linked data points in a single experiment.
- Feeding these data into protein‑language AI models enabled rapid, accurate prediction of beneficial mutations for several enzymes, including a miniature CRISPR‑Cas protein.
- The approach couples experiment and computation, reducing the protein‑optimization cycle from months to just three days in proof‑of‑concept studies.
- Funding came from multiple sources, including the NIH, Welch Foundation, DOD, and private foundations, underscoring broad interest in AI‑driven protein design.
The Challenge of Vast Sequence Space
Protein engineering seeks to tweak amino‑acid sequences to improve function, yet the combinatorial explosion makes exhaustive testing impossible. As the article notes, “For a protein that is just 50 amino acids in length, this leads to approximately 1.13×10⁶⁵ potential combinations to test – that’s 113 followed by 65 zeros, or five times as many zeros as a trillion has.” This astronomical number renders traditional trial‑and‑error impractical, positioning AI as a natural candidate to navigate the landscape. However, the utility of any machine‑learning model rests on the quality and quantity of data used to train it—a limitation that has plagued many protein‑engineering projects.
Data Scarcity as the Core Bottleneck
While constructing sophisticated algorithms is relatively straightforward, obtaining enough experimental measurements to feed those models remains the true obstacle. Han Xiao, professor of chemistry, biosciences and bioengineering at Rice University and director of the SynthX Center, emphasized this point: “One of the biggest bottlenecks in AI-guided protein engineering is not coming up with machine-learning models. It is generating the right and enough experimental data to train them.” For activity‑focused engineering—optimizing what a protein actually does—publicly available datasets were simply insufficient, hindering the development of predictive models.
Enter Sequence Display: A High‑Throughput Barcoding Strategy
To overcome the data deficit, Xiao’s team devised Sequence Display, a method that couples mutagenesis with a programmable barcode system. As graduate student Linqi Cheng, the study’s first author, explained, “We were able to develop an activity‑based barcoding system that records the activity of individual protein variants and generates the kind of dataset needed to train a machine learning model.” In practice, each variant of a target protein receives a unique DNA barcode. A special editor molecule modifies that barcode in proportion to the protein’s activity level, so more active variants acquire larger barcode changes. After the reaction, next‑generation sequencing reads the barcodes, translating sequencing read counts into quantitative activity scores for millions of variants in a single assay.
From Barcodes to Millions of Data Points
The power of Sequence Display lies in its scale. A single experiment can yield “more than 10 million data points,” providing a rich, activity‑annotated sequence landscape that AI models can consume. These data are then fed into protein‑language models—neural networks trained to understand the statistical relationships between amino‑acid sequences and functional outcomes. By learning from the empirical activity distribution, the models can predict which mutations are likely to enhance or diminish a protein’s performance, effectively narrowing the astronomical search space to a manageable set of promising candidates.
Proof‑of‑Concept with a Miniature CRISPR‑Cas Protein
To validate the workflow, the researchers selected a small CRISPR‑Cas protein valued for its compact size but limited DNA‑targeting breadth. They mutagenized the gene encoding this Cas variant, attached the activity‑responsive barcodes, and ran the Sequence Display assay. The resulting dataset captured a fine‑grained activity profile across millions of mutants. Cheng highlighted the synergy: “The AI is not replacing the experiment here. It instead depends on the experiment… Sequence Display gives us the data foundation, and the models help us search a much larger data space for strong candidates.” Using the model’s predictions, they identified mutations that broadened the protein’s DNA‑recognition scope, demonstrating that activity‑guided AI design can yield functional improvements in days rather than months.
Expanding the Approach to Other Enzymes
Encouraged by the initial success, the team applied Sequence Display to additional proteins, including aminoacyl‑tRNA synthetases, cytosine deaminase, and uracil glycosylase inhibitor. In each case, the barcoding experiment produced sufficient activity data to train accurate AI models, which then suggested mutations that significantly boosted enzymatic performance. This reproducibility underscores the method’s generality: any protein whose activity can be linked to a measurable signal—fluorescence, absorbance, or a sequencing‑readable barcode—can be subjected to the same pipeline.
A Framework for Integrated AI‑Driven Protein Engineering
Han Xiao summarized the broader implication of the work: “What this approach provides is a practical framework for integrating AI with protein engineering. Rather than relying on machine learning as a stand-alone solution, we couple it with an experimental platform that generates high-quality training data. This synergy enables more efficient discovery of advanced research tools and next‑generation therapeutic proteins.” By closing the loop between experiment and prediction, Sequence Display transforms AI from a speculative tool into a reliable engine for protein optimization, accelerating the development of reagents for basic research, diagnostics, and therapeutics.
Funding and Acknowledgments
The study received support from a variety of sources, reflecting the interdisciplinary and translational nature of the project. Notable contributors include a SynthX Seed Award (SYN‑IN‑2024‑002), multiple NIH grants (R35‑GM133706, R01‑CA277838, R01‑AI165079), the Robert A. Welch Foundation (C‑1970), the U.S. Department of Defense (W81XWH‑21‑1‑0789, HT9425‑23‑1‑0494, HT9425‑25‑1‑0021), a Rice Synthetic Biology Institute Seed Grant, and a Medical Research Award from the Robert J. Kleberg, Jr. and Helen C. Kleberg Foundation. The findings were published in Nature Biotechnology (Cheng L. et al., 2026, DOI: 10.1038/s41587‑026‑03087‑3).
https://www.news-medical.net/news/20260413/New-method-boosts-AI-driven-protein-engineering-with-massive-data.aspx